Anhui University

Yingjian Deng

Jie Chen

Zihan Cheng

Bing Li

Haitao Wang

Jingmin Xi

The present disclosure provides a method for fine-grained detection of driver distraction based on unsupervised learning, belonging to the field of driving behavior analysis. The method includes: acquiring distracted driving image data; and inputting the acquired distracted driving image data into an unsupervised learning detection model, analyzing the distracted driving image data by using the unsupervised learning detection model, and determining a driver distraction state according to an analysis result. The unsupervised learning detection model includes a backbone network, projection heads, and a loss function; the backbone network is a RepMLP network structure incorporating a multilayer perceptron (MLP); the projection heads are each an MLP incorporating a residual structure; and the loss function is a loss function based on contrastive learning and a stop-gradient strategy.

Method for fine-grained detection of driver distraction based on unsupervised learning

The present disclosure provides a transformer-based driver distraction detection method and apparatus, belonging to the field of driving behavior analysis. The method includes: acquiring districted driving image data; building a driver distraction detection model FPT; inputting the acquired distracted driving image data into the driver distraction detection model FPT, analyzing the distracted driving image data by using the driver distraction detection model FPT, and determining a driver distraction state according to an analysis result. The present disclosure proposes a new network model, i.e., a driver distraction detection model FPT, based on Swin, Twins, and other models. Compared with a deep learning model, the FPT model compensates for the drawback that the deep learning model can only extract local features; the FPT model improves the classification accuracy and reduces the parameter quantity and calculation amount compared with the transformer model. The present disclosure adjusts the loss function of the whole network and adds label smoothing to the cross-entropy loss function, to increase the accuracy of classification, effectively suppress overfitting, and improve the detection accuracy.

Growing community of inventors

Hefei, China

Yingjian Deng